April 21, 2019

Introduction

Transit method is used to detect the presence of exoplanets around distant stars. Here’s a helpful visualization:

Data

  • Data was obtained from the Kepler space telescope. Data format: time series (equal intervals) with some missing data
  • Only the light flux was used.
  • Data is not as clean as expected.
    • Universe is large - stars exhibit variable brightness at large distances due to several phenomenon.
    • Not all crests and troughs are relevant.

Components

LSTM autoencoder hyperparameters:

  • Train-validation-test: 42%, 28%, 30%
  • Sequence length = 10
  • Layers: input (9, 1), LSTM (64), droupout (0.2), LSTM (256), droupout (0.2), LSTM (100), droupout (0.2), dense (9, 1)
  • Loss: Mean square error

Results: Learning and Testing

Scope for Improvement

  • Hyperparameters such as sequence length, number of hidden units in LSTM layers, number of dense layers, number of neurons in dense layer, etc. can be tuned.
  • Bi-directional LSTM autoencoder can be used instead of directional LSTM autoencoder.
  • An algorithm can be used to perform phase folding to identify number of planets and period of revolution of each planet around the star. This will also enrich the data.